Emergency control strategy model of power system extreme disturbance based on spatial–temporal graph attention network

As the share of new energy in power grids rises, power electronic equipment penetration increases, raising the probability of extreme disturbances. Existing emergency control technologies have clear limitations. New post-2023 methods like STGCN + PPO and Transformer + DRL inaccurately capture spatiotemporal features, struggle to balance control accuracy and real-time performance, and fail to meet high-renewable grid requirements. This study addresses these issues using a Spatial Temporal Graph Attention Network (STGAT) for extreme disturbance control. We construct an integrated framework for adaptive feature extraction and multi-objective decision-making, incorporating a spatiotemporal attention module, multi-source data interaction, and control deviation calibration. Key innovations include dynamic attention weight allocation, reduced feature transfer loss, and disturbance intensity classification. Experiments show the STGAT model achieves a comprehensive score of 0.94–30.6% higher than traditional GCN. Disturbance recognition accuracy reaches 98.7% (6.7% above STGCN), control error drops to 2.1% (64.4% lower than DDQN), robustness coefficient is 0.15 (48.3% better than GAT + DDQN), and practical adaptability is 0.95. All indices surpass post-2023 benchmarks. This technology accurately captures extreme disturbance propagation, generates optimal control strategies, and enhances system anti-disturbance capability, offering significant engineering value for high-renewable grids.